1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright (C) 2015 Davidlohr Bueso.
5 * Block a bunch of threads and let parallel waker threads wakeup an
6 * equal amount of them. The program output reflects the avg latency
7 * for each individual thread to service its share of work. Ultimately
8 * it can be used to measure futex_wake() changes.
11 #include <linux/compiler.h>
12 #include "../util/debug.h"
14 #ifndef HAVE_PTHREAD_BARRIER
15 int bench_futex_wake_parallel(int argc __maybe_unused, const char **argv __maybe_unused)
17 pr_err("%s: pthread_barrier_t unavailable, disabling this test...\n", __func__);
20 #else /* HAVE_PTHREAD_BARRIER */
21 /* For the CLR_() macros */
26 #include "../util/stat.h"
27 #include <subcmd/parse-options.h>
28 #include <linux/kernel.h>
29 #include <linux/time64.h>
32 #include <internal/cpumap.h>
33 #include <perf/cpumap.h>
42 struct timeval runtime;
45 static unsigned int nwakes = 1;
47 /* all threads will block on the same futex -- hash bucket chaos ;) */
48 static u_int32_t futex = 0;
50 static pthread_t *blocked_worker;
51 static bool done = false, silent = false, fshared = false;
52 static unsigned int nblocked_threads = 0, nwaking_threads = 0;
53 static pthread_mutex_t thread_lock;
54 static pthread_cond_t thread_parent, thread_worker;
55 static pthread_barrier_t barrier;
56 static struct stats waketime_stats, wakeup_stats;
57 static unsigned int threads_starting;
58 static int futex_flag = 0;
60 static const struct option options[] = {
61 OPT_UINTEGER('t', "threads", &nblocked_threads, "Specify amount of threads"),
62 OPT_UINTEGER('w', "nwakers", &nwaking_threads, "Specify amount of waking threads"),
63 OPT_BOOLEAN( 's', "silent", &silent, "Silent mode: do not display data/details"),
64 OPT_BOOLEAN( 'S', "shared", &fshared, "Use shared futexes instead of private ones"),
68 static const char * const bench_futex_wake_parallel_usage[] = {
69 "perf bench futex wake-parallel <options>",
73 static void *waking_workerfn(void *arg)
75 struct thread_data *waker = (struct thread_data *) arg;
76 struct timeval start, end;
78 pthread_barrier_wait(&barrier);
80 gettimeofday(&start, NULL);
82 waker->nwoken = futex_wake(&futex, nwakes, futex_flag);
83 if (waker->nwoken != nwakes)
84 warnx("couldn't wakeup all tasks (%d/%d)",
85 waker->nwoken, nwakes);
87 gettimeofday(&end, NULL);
88 timersub(&end, &start, &waker->runtime);
94 static void wakeup_threads(struct thread_data *td, pthread_attr_t thread_attr)
98 pthread_attr_setdetachstate(&thread_attr, PTHREAD_CREATE_JOINABLE);
100 pthread_barrier_init(&barrier, NULL, nwaking_threads + 1);
102 /* create and block all threads */
103 for (i = 0; i < nwaking_threads; i++) {
105 * Thread creation order will impact per-thread latency
106 * as it will affect the order to acquire the hb spinlock.
107 * For now let the scheduler decide.
109 if (pthread_create(&td[i].worker, &thread_attr,
110 waking_workerfn, (void *)&td[i]))
111 err(EXIT_FAILURE, "pthread_create");
114 pthread_barrier_wait(&barrier);
116 for (i = 0; i < nwaking_threads; i++)
117 if (pthread_join(td[i].worker, NULL))
118 err(EXIT_FAILURE, "pthread_join");
120 pthread_barrier_destroy(&barrier);
123 static void *blocked_workerfn(void *arg __maybe_unused)
125 pthread_mutex_lock(&thread_lock);
127 if (!threads_starting)
128 pthread_cond_signal(&thread_parent);
129 pthread_cond_wait(&thread_worker, &thread_lock);
130 pthread_mutex_unlock(&thread_lock);
132 while (1) { /* handle spurious wakeups */
133 if (futex_wait(&futex, 0, NULL, futex_flag) != EINTR)
141 static void block_threads(pthread_t *w, pthread_attr_t thread_attr,
142 struct perf_cpu_map *cpu)
147 threads_starting = nblocked_threads;
149 /* create and block all threads */
150 for (i = 0; i < nblocked_threads; i++) {
152 CPU_SET(cpu->map[i % cpu->nr], &cpuset);
154 if (pthread_attr_setaffinity_np(&thread_attr, sizeof(cpu_set_t), &cpuset))
155 err(EXIT_FAILURE, "pthread_attr_setaffinity_np");
157 if (pthread_create(&w[i], &thread_attr, blocked_workerfn, NULL))
158 err(EXIT_FAILURE, "pthread_create");
162 static void print_run(struct thread_data *waking_worker, unsigned int run_num)
164 unsigned int i, wakeup_avg;
165 double waketime_avg, waketime_stddev;
166 struct stats __waketime_stats, __wakeup_stats;
168 init_stats(&__wakeup_stats);
169 init_stats(&__waketime_stats);
171 for (i = 0; i < nwaking_threads; i++) {
172 update_stats(&__waketime_stats, waking_worker[i].runtime.tv_usec);
173 update_stats(&__wakeup_stats, waking_worker[i].nwoken);
176 waketime_avg = avg_stats(&__waketime_stats);
177 waketime_stddev = stddev_stats(&__waketime_stats);
178 wakeup_avg = avg_stats(&__wakeup_stats);
180 printf("[Run %d]: Avg per-thread latency (waking %d/%d threads) "
181 "in %.4f ms (+-%.2f%%)\n", run_num + 1, wakeup_avg,
182 nblocked_threads, waketime_avg / USEC_PER_MSEC,
183 rel_stddev_stats(waketime_stddev, waketime_avg));
186 static void print_summary(void)
188 unsigned int wakeup_avg;
189 double waketime_avg, waketime_stddev;
191 waketime_avg = avg_stats(&waketime_stats);
192 waketime_stddev = stddev_stats(&waketime_stats);
193 wakeup_avg = avg_stats(&wakeup_stats);
195 printf("Avg per-thread latency (waking %d/%d threads) in %.4f ms (+-%.2f%%)\n",
198 waketime_avg / USEC_PER_MSEC,
199 rel_stddev_stats(waketime_stddev, waketime_avg));
203 static void do_run_stats(struct thread_data *waking_worker)
207 for (i = 0; i < nwaking_threads; i++) {
208 update_stats(&waketime_stats, waking_worker[i].runtime.tv_usec);
209 update_stats(&wakeup_stats, waking_worker[i].nwoken);
214 static void toggle_done(int sig __maybe_unused,
215 siginfo_t *info __maybe_unused,
216 void *uc __maybe_unused)
221 int bench_futex_wake_parallel(int argc, const char **argv)
225 struct sigaction act;
226 pthread_attr_t thread_attr;
227 struct thread_data *waking_worker;
228 struct perf_cpu_map *cpu;
230 argc = parse_options(argc, argv, options,
231 bench_futex_wake_parallel_usage, 0);
233 usage_with_options(bench_futex_wake_parallel_usage, options);
237 sigfillset(&act.sa_mask);
238 act.sa_sigaction = toggle_done;
239 sigaction(SIGINT, &act, NULL);
241 cpu = perf_cpu_map__new(NULL);
243 err(EXIT_FAILURE, "calloc");
245 if (!nblocked_threads)
246 nblocked_threads = cpu->nr;
248 /* some sanity checks */
249 if (nwaking_threads > nblocked_threads || !nwaking_threads)
250 nwaking_threads = nblocked_threads;
252 if (nblocked_threads % nwaking_threads)
253 errx(EXIT_FAILURE, "Must be perfectly divisible");
255 * Each thread will wakeup nwakes tasks in
256 * a single futex_wait call.
258 nwakes = nblocked_threads/nwaking_threads;
260 blocked_worker = calloc(nblocked_threads, sizeof(*blocked_worker));
262 err(EXIT_FAILURE, "calloc");
265 futex_flag = FUTEX_PRIVATE_FLAG;
267 printf("Run summary [PID %d]: blocking on %d threads (at [%s] "
268 "futex %p), %d threads waking up %d at a time.\n\n",
269 getpid(), nblocked_threads, fshared ? "shared":"private",
270 &futex, nwaking_threads, nwakes);
272 init_stats(&wakeup_stats);
273 init_stats(&waketime_stats);
275 pthread_attr_init(&thread_attr);
276 pthread_mutex_init(&thread_lock, NULL);
277 pthread_cond_init(&thread_parent, NULL);
278 pthread_cond_init(&thread_worker, NULL);
280 for (j = 0; j < bench_repeat && !done; j++) {
281 waking_worker = calloc(nwaking_threads, sizeof(*waking_worker));
283 err(EXIT_FAILURE, "calloc");
285 /* create, launch & block all threads */
286 block_threads(blocked_worker, thread_attr, cpu);
288 /* make sure all threads are already blocked */
289 pthread_mutex_lock(&thread_lock);
290 while (threads_starting)
291 pthread_cond_wait(&thread_parent, &thread_lock);
292 pthread_cond_broadcast(&thread_worker);
293 pthread_mutex_unlock(&thread_lock);
297 /* Ok, all threads are patiently blocked, start waking folks up */
298 wakeup_threads(waking_worker, thread_attr);
300 for (i = 0; i < nblocked_threads; i++) {
301 ret = pthread_join(blocked_worker[i], NULL);
303 err(EXIT_FAILURE, "pthread_join");
306 do_run_stats(waking_worker);
308 print_run(waking_worker, j);
313 /* cleanup & report results */
314 pthread_cond_destroy(&thread_parent);
315 pthread_cond_destroy(&thread_worker);
316 pthread_mutex_destroy(&thread_lock);
317 pthread_attr_destroy(&thread_attr);
321 free(blocked_worker);
324 #endif /* HAVE_PTHREAD_BARRIER */